GHSA-j8f7-x8jm-wmm4: Langflow: SSRF in built-in components exposes cloud metadata

GHSA-j8f7-x8jm-wmm4MEDIUM
Published October 6, 2026
CISO Take

Before Langflow 1.10.3, several built-in components (RSS Reader, SearXNG, Web Search, Home Assistant, Glean, Docling Serve) fetched author-controlled URLs without SSRF checks. Any authenticated user who can build or run flows could make the Langflow server request internal networks or cloud metadata endpoints such as 169.254.169.254, and often read the response in the component output. That can leak IAM credentials and open a path to lateral movement. Exposure is limited by the authentication requirement. There is no KEV listing, no public exploit and no Nuclei template. EPSS and CVSS are not yet scored, and the advisory is rated medium. Langflow is widely used for building agentic workflows, and its package risk score is 77/100 with 194 other CVEs on record. Upgrade to langflow 1.10.3 (langflow-base 0.10.3, lfx 1.10.3, lfx-docling 0.1.2). Keep LANGFLOW_SSRF_PROTECTION_ENABLED at its default of true. Where you cannot patch, enforce IMDSv2 and block egress to link-local and private ranges.

Sources:GitHub AdvisoryATLAS

What is the risk?

Medium overall, but higher for multi-tenant or shared Langflow instances on cloud VMs. Exploitation is trivial: set a component URL to a metadata or internal address and run the flow. It requires an authenticated account with flow-building rights, so deployments where every user is trusted face lower risk. The impact includes credential theft from cloud metadata and internal service reconnaissance. Mitigating signals: no KEV listing, no known public exploit, no scanner template. Gaps in the data: EPSS and CVSS are not yet scored. Releases 1.7.0 to 1.9.2 had the guard disabled by default, and 1.9.3 to 1.10.2 still left some components unprotected, so partially updated installs remain exposed.

How does the attack unfold?

Initial Access
Attacker logs in with a valid account that can build or run flows on the Langflow server.
AML.T0012
Exploitation
Attacker sets an RSS Reader, SearXNG, Home Assistant, Glean or Docling Serve URL to an internal or metadata address and runs the flow.
AML.T0049
Credential Access
The server fetches the target and returns the response, exposing cloud IAM credentials or internal service data.
AML.T0055
Impact
Stolen credentials are used to reach cloud storage, model artifacts and internal AI infrastructure.

What systems are affected?

PackageEcosystemVulnerable RangePatched
Langflowpip< 1.10.31.10.3
156.0KPushed 6d ago18% patched~66d to patchFull package profile →
Langflowpip< 0.10.30.10.3
156.0KPushed 6d ago18% patched~66d to patchFull package profile →
Langflowpip< 1.10.31.10.3
156.0KPushed 6d ago18% patched~66d to patchFull package profile →
Langflowpip< 0.1.20.1.2
156.0KPushed 6d ago18% patched~66d to patchFull package profile →

How severe is it?

CVSS 3.1
N/A
EPSS
N/A
Exploitation Status
No known exploitation
Sophistication
Trivial

What should I do?

1 step
  1. 1) Upgrade to langflow >= 1.10.3 (langflow-base >= 0.10.3, lfx >= 1.10.3, lfx-docling >= 0.1.2). 2) Do not set LANGFLOW_SSRF_PROTECTION_ENABLED=false. If internal hosts must be reachable, allowlist only those via LANGFLOW_SSRF_ALLOWED_HOSTS. Review LANGFLOW_CONNECTOR_SSRF_VALIDATION_ENABLED for connector components. 3) If you cannot patch, block egress from the Langflow host to 169.254.169.254, loopback and RFC 1918 ranges, and enforce IMDSv2 with a hop limit of 1 on AWS. 4) Restrict flow-building permissions to trusted users and review existing flows for components pointing at internal URLs. 5) Detect: alert on Langflow host requests to metadata IPs, review cloud audit logs for metadata-credential use from unexpected IPs, and search logs for SSRFProtectionError after upgrading.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Art. 15-Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MANAGE 2.3-Responses to previously unknown risks
OWASP LLM Top 10
LLM03-Supply ChainLLM06-Excessive Agency

Frequently Asked Questions

What is GHSA-j8f7-x8jm-wmm4?

Before Langflow 1.10.3, several built-in components (RSS Reader, SearXNG, Web Search, Home Assistant, Glean, Docling Serve) fetched author-controlled URLs without SSRF checks. Any authenticated user who can build or run flows could make the Langflow server request internal networks or cloud metadata endpoints such as 169.254.169.254, and often read the response in the component output. That can leak IAM credentials and open a path to lateral movement. Exposure is limited by the authentication requirement. There is no KEV listing, no public exploit and no Nuclei template. EPSS and CVSS are not yet scored, and the advisory is rated medium. Langflow is widely used for building agentic workflows, and its package risk score is 77/100 with 194 other CVEs on record. Upgrade to langflow 1.10.3 (langflow-base 0.10.3, lfx 1.10.3, lfx-docling 0.1.2). Keep LANGFLOW_SSRF_PROTECTION_ENABLED at its default of true. Where you cannot patch, enforce IMDSv2 and block egress to link-local and private ranges.

Is GHSA-j8f7-x8jm-wmm4 actively exploited?

No confirmed active exploitation of GHSA-j8f7-x8jm-wmm4 has been reported, but organizations should still patch proactively.

How to fix GHSA-j8f7-x8jm-wmm4?

1) Upgrade to langflow >= 1.10.3 (langflow-base >= 0.10.3, lfx >= 1.10.3, lfx-docling >= 0.1.2). 2) Do not set LANGFLOW_SSRF_PROTECTION_ENABLED=false. If internal hosts must be reachable, allowlist only those via LANGFLOW_SSRF_ALLOWED_HOSTS. Review LANGFLOW_CONNECTOR_SSRF_VALIDATION_ENABLED for connector components. 3) If you cannot patch, block egress from the Langflow host to 169.254.169.254, loopback and RFC 1918 ranges, and enforce IMDSv2 with a hop limit of 1 on AWS. 4) Restrict flow-building permissions to trusted users and review existing flows for components pointing at internal URLs. 5) Detect: alert on Langflow host requests to metadata IPs, review cloud audit logs for metadata-credential use from unexpected IPs, and search logs for SSRFProtectionError after upgrading.

What systems are affected by GHSA-j8f7-x8jm-wmm4?

This vulnerability affects the following AI/ML architecture patterns: Agent frameworks, LLM workflow orchestration, RAG pipelines, Low-code AI builders, Cloud-hosted AI platforms.

What is the CVSS score for GHSA-j8f7-x8jm-wmm4?

No CVSS score has been assigned yet.

What is the AI security impact?

Affected AI Architectures

Agent frameworksLLM workflow orchestrationRAG pipelinesLow-code AI buildersCloud-hosted AI platforms

MITRE ATLAS Techniques

AML.T0012Valid Accounts
AML.T0049Exploit Public-Facing Application
AML.T0055Unsecured Credentials
AML.T0075Cloud Service Discovery

Compliance Controls Affected

EU AI Act: Art. 15
ISO 42001: A.6.2.6
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM03, LLM06

What are the technical details?

Original Advisory

### Summary Before **Langflow 1.10.3**, several built-in components sent server-side HTTP requests to URLs the flow author controls, with no SSRF enforcement. An authenticated user who can build or run flows could make the Langflow server reach loopback, RFC 1918 / private networks, link-local addresses and cloud metadata endpoints (for example `http://169.254.169.254/latest/meta-data/`), and in many cases read the response back through the component output. Two things combined to cause this: 1. **Protection disabled / warn-only.** The SSRF guard (`lfx/utils/ssrf_protection.py`, added in 1.7.0 by #10544) shipped with `ssrf_protection_enabled = False`, and its only caller, the **API Request** component, called `validate_url_for_ssrf(url, warn_only=True)`. So even with protection turned on, blocked URLs were only logged. 2. **Inconsistent coverage.** Other components that take a user-supplied URL or base URL did not call the guard at all: RSS Reader, SearXNG, Web Search (fetching result / RSS URLs), Home Assistant, Glean Search and Docling Serve (remote). The problem was fixed across several PRs and is fully fixed in **Langflow 1.10.3** (and 1.11.0+). With the default settings, outbound requests from these components are now validated after DNS resolution, pinned to the validated IP (to prevent DNS rebinding) and **blocked** when they target private, loopback, link-local or metadata addresses, unless the operator allowlists the host. ### Affected versions | Package (PyPI) | Vulnerable | Patched | |---|---|---| | `langflow` | `< 1.10.3` | `1.10.3` | | `langflow-base` | `< 0.10.3` | `0.10.3` | | `lfx` | `< 1.10.3` | `1.10.3` | | `lfx-docling` (Docling Serve component) | `< 0.1.2` | `0.1.2` | How the fix rolled out: | Release | Change | |---|---| | `< 1.7.0` | No SSRF guard. All URL-taking components make unrestricted requests. | | `1.7.0` – `1.9.2` | Guard exists, but `ssrf_protection_enabled=False` by default and API Request uses `warn_only=True`. | | `1.9.3` | #13016: protection **on by default**, API Request **blocks** (DNS pinning / anti-rebinding). Other components still unprotected. | | `1.10.1` | #13395 (RSS Reader, SearXNG) and #13643 (Home Assistant, Glean, plus model-provider URL components) route requests through SSRF-safe helpers. | | `1.10.3` | #14033 (Docling Serve) and #14071 (Web Search result/RSS fetching). **All components cited in this report are protected.** | ### Details Vulnerable pattern (for example `rss.py` before #13395): ```python response = requests.get(self.rss_url, timeout=self.timeout) ``` API Request before #13016: ```python # TODO: In next major version (2.0), remove warn_only=True to enforce blocking validate_url_for_ssrf(url, warn_only=True) ``` with `ssrf_protection_enabled: bool = False` in the settings. **Clarifications on the original report.** Some sinks in the report are not exploitable SSRF: - `web_search.py`: the DuckDuckGo search request and the Google News RSS request, and `news_search.py:131`. The host is hard-coded (`html.duckduckgo.com` / `news.google.com`) and user input only goes into `quote_plus`-encoded path and query segments, so the destination cannot be redirected. The real Web Search sink was fetching the **result URLs**, which is fixed in #14071. - `langflow/load/utils.py` `get_flow()`: this is a client-side SDK helper (`from langflow.load import get_flow`) that the caller runs against their own Langflow URL. No API route calls it, so it is not a backend SSRF endpoint. ### PoC (vulnerable versions) 1. Log in as any user who can create flows. 2. Add an **RSS Reader** (or SearXNG / Home Assistant / Glean / Docling Serve) component and set its URL to `http://169.254.169.254/latest/meta-data/` (or an internal host such as `http://10.0.0.1/`). 3. Run the component. The Langflow server issues the request, and the response or error is shown in the component output. On patched versions the same run fails with `SSRFProtectionError: Access to IP address 169.254.169.254 is blocked by SSRF protection`. ### Impact - Access to cloud instance metadata (IAM credentials / tokens) on AWS, GCP and Azure deployments. - Probing and interaction with internal services and private networks reachable from the Langflow host. - Requires an authenticated account with permission to build or run flows. Deployments where every user is already fully trusted are less affected. ### Fix - **#13016**: sets `ssrf_protection_enabled = True` by default, changes `warn_only` to default `False`, drops `warn_only=True` from API Request, and adds `validate_and_resolve_url()` plus `SSRFProtectedTransport` (DNS pinning). ```diff - ssrf_protection_enabled: bool = False + ssrf_protection_enabled: bool = True ``` - **#13395**: adds `lfx.utils.ssrf_requests.ssrf_safe_get` (per-hop validation and pinning for `requests`) and uses it in RSS Reader and SearXNG. ```diff - response = requests.get(self.rss_url, timeout=self.timeout) + response = ssrf_safe_get(self.rss_url, timeout=self.timeout) ``` - **#13643**: adds `lfx.utils.ssrf_httpx` (`ssrf_safe_httpx_get/post`, protected client kwargs) and applies it to Home Assistant, Glean and the model-provider base-URL components. - **#14033**: Docling Serve builds its `httpx.Client` with `ssrf_protected_httpx_client_kwargs_for_url(base_url)`. - **#14071** (release-1.10.3) / **#14073** (forward-port to 1.11.0): Web Search validates and pins each result URL before fetching it, with redirects disabled. Note: connector-style components (for example Home Assistant, Ollama, LM Studio) keep a documented loopback exemption (#14146), so a self-hosted user can reach a local service on `localhost`. Private-network and metadata addresses stay blocked. Operators can tighten or relax this with `LANGFLOW_SSRF_ALLOWED_HOSTS` and `LANGFLOW_CONNECTOR_SSRF_VALIDATION_ENABLED`. ### Remediation - Upgrade to `langflow >= 1.10.3` (or `langflow-base >= 0.10.3` / `lfx >= 1.10.3` / `lfx-docling >= 0.1.2`). - Do not set `LANGFLOW_SSRF_PROTECTION_ENABLED=false`. If you need to reach internal hosts on purpose, allowlist only those hosts with `LANGFLOW_SSRF_ALLOWED_HOSTS`. - On versions that cannot be upgraded right away, restrict outbound network access from the Langflow host (for example, block `169.254.169.254` and use IMDSv2 on AWS).

Exploitation Scenario

An attacker obtains a low-privilege Langflow account (a compromised developer login or a shared tenant). They create a flow with an RSS Reader component whose URL is http://169.254.169.254/latest/meta-data/iam/security-credentials/. They run the flow and read the role name and temporary credentials in the component output. They then use those credentials from outside the network to access S3 buckets holding training data, model artifacts or vector store snapshots. They can also point other components at internal admin panels or APIs to map the environment.

Weaknesses (CWE)

CWE-918 — Server-Side Request Forgery (SSRF): The web server receives a URL or similar request from an upstream component and retrieves the contents of this URL, but it does not sufficiently ensure that the request is being sent to the expected destination.

Source: MITRE CWE corpus.

Timeline

Published
October 6, 2026
Last Modified
October 6, 2026
First Seen
October 6, 2026

Related Vulnerabilities